Clarify QSPR/ChemProp scaling for “window” vs modifiers: do model outputs need pre-scaling?

[mmahmedgithub]

Context

I’m using DrugEx Graph RL with multiple objectives:

• QSPR/ChemProp regressors: absolute LogP, absolute LogD7, log‑scaled Caco2
• RDKit properties: QED, MW, TPSA, logP

DrugEx applies modifiers (e.g., ClippedScore, SmoothHump) to map raw objective values to [0,1], then combines them via WS or Pareto.

---

Observed Behavior

For QSPR “window” targets, the modifier requires the targets to be 0‑centered (roughly ±1.5 around 0, hard coded it seems).
This works if the model output (my qsprpred/chemprop) is a residual or zero‑centered value, but fails for absolute outputs like LogP ≈ 2–5 (scores collapse toward 0 unless pre‑centered).

RDKit windows accept explicit low/high thresholds and work fine for absolute ranges.

---

Questions

1. For QSPR/ChemProp regressors on absolute scales (LogP, LogD7, log‑Caco2),
   do outputs need to be pre‑standardized/zero‑centered to use “window”? . This eventually require me retraining the models after rescaling the endpoints to the desired 0-centered residuals, am I getting things right here ?
   Or can modifiers map absolute values directly to [0,1] without retraining?

2. Is there a supported way to specify an arbitrary window (low, high, sharpness) for QSPR objectives (unscaled/real values, like logp, pic50, pki, and so on), similar to RDKit windows?

3. If not, is this the recommended pattern?

   • Use QSPR active/inactive with thresholds for absolute regressors. But even here, it seems that I need to specify a single cut-off / threshold that will be used for all endpoints (not sure if this will work for scaled outputs). If you are familiar with reinvent4, may be you can get where does the confusion come from.
   • Keep range constraints via RDKit property windows

4. For WS vs Pareto:

   • Are per‑objective pass thresholds (e.g., ~0.99) the only condition for Desired=1?
   • Do modifiers ensure scale invariance across heterogeneous objectives?

5. Are there minimal example configs combining:

   • QSPR absolute outputs (LogP/LogD)
   • QSPR “low‑is‑better” (Caco2) . Yes, this is what i need in my specific problem, low Caco2
   • RDKit windows
     under a single reward scheme without pre‑transforming model outputs?

Environment

• DrugEx: master (containerized)
• CUDA: RTX 4070S
• ChemProp models: standard regression (absolute) and log‑scaled Caco2

---

Goal

Confirm whether QSPR “window” is intentionally zero‑centered only, and whether there could be an option to set (low, high, shape) directly for QSPR window objectives to avoid retraining models to residuals.

[HellevdM]

Hi @mmahmedgithub,

(1) The QSPR window targets of the command line interface were implemented for this paper Schoenmaker, L., Béquignon, O.J.M., Jespers, W. et al. UnCorrupt SMILES: a novel approach to de novo design. These "windows" were specifically used for predicting the selectivity of one target over another. I think in your use case it is not really applicable.

(2+ 3) In general, the command line interface is limited regarding the specification of modifiers. If you need more flexibility you could use the Python API to specify all parameters of the modifiers or add an additional modifier yourself. There is an example in the sequence RNN tutorial of setting a ClippedScore and SmoothClippedScore modifier. Let us know if that answers your question, if not it would be very helpful if you could go into the details of your use case a bit further (i.e., what do the predictions look like, what do you want to achieve)

(4)

• In addition to the per-objective thresholds, there is also the requirement that the generated molecule is valid (i.e. that it can be parsed by the RDkit).
• Most of the modifier functions should scale objectives between 0-1. But it is up to the user to check that the selected modifier and settings make sense for their target. For example if your objective has values very close to zero, the default clipped score modifier settings will not give adequate scaling of the objective values.

(5) We don't have any minimal examples available for these cases. For the "low-is-better" case, the values will have to be inverted as the reward schemes are defined for the "high-is-better" case.

I hope this answers your questions!

Best,
Helle

[mmahmedgithub]

Thanks, Helle — that clears up a lot. I still have a few ambiguities and a concrete use case to sanity-check.

My use case (goal-oriented, multi-parameter optimization of a given hit/scaffold)

*Hit-to-lead/lead-opt around fixed scaffolds with mixed objectives:

*Improve target-A potency: pIC50_A ~6 → ≥8

*Reduce hERG liability: pIC50_hERG ~6 → ≤4

*Improve solubility: logS ~−4 → ≥−3

*Keep physchem in windows: logP ∈ [2,4], HBD ∈ [1,5], HBA ∈ [1,10], RB ∈ [2,8], etc.
Inputs are a set of QSPR regressors (ChemProp/qsprpred, absolute scales for LogP, LogD7,herg, log-Caco2, pIC50/pKi) plus RDKit properties. DrugEx Graph RL is asked to generate molecules satisfying these predicted targets.

Follow-ups

1 - QSPR “window” behavior
To confirm: the CLI “QSPR window” is intentionally for zero-centered/selectivity use (per UnCorrupt SMILES) and not for absolute ranges. If so, what other CLI/API modules/modifier can support an absolute-value window for QSPR objectives (parameters: low, high, optional sharpness) so we don’t have to retrain models to zero-centered residuals?

2 - Python API modifiers for absolute outputs
If the recommended path is the Python API, could you share (or point me to) a tiny snippet showing how to attach different modifiers per objective on their native scales? Something along the lines of:

smooth “high-is-better” band for pIC50_A (peaks above ~8)

inverted/sigmoid “low-is-better” for pIC50_hERG (falls below ~4)

hump/window for logS (target ≥−3, taper outside)

Any other regressors returning floats, classifiers returning probabilities, rdkit properties windows, and so on. They do not have to be the exact same properties I listed here or in the old thread, any endpoints of heterogenous nature are fine

[martin-sicho]

Hi @mmahmedgithub,

Regarding your second point, here are some examples for the use cases that you are mentioning. You can use the modifier Python API to implement those:

This is the SmoothClippedScore modifier that you can apply to any continuous value. In this case, we are setting it to score any values lower than -4 as ideal (-infinity maps to 1.0). Naturally, you can define the sigmoid also in the opposite direction (higher is better):

Here all values above 8 are ideal and +infinity maps to 1.0.

In addition, we have the SmoothHump function which you can use to define a range:

This is an example with a peak ideal value around -3.

Regarding your first point about windows, you can easily implement your own modifiers with the Python API. Just subclass ScoreModifier and add it to your workflow. For example, I can imagine that you could combine several SmoothHump functions to get center value peaks for the window intervals you require. You will find some Python code that will allow you to play with the modifiers in this tutorial. FYI: The images I showed are from the frontend of the GenUI app that we are also developing so that can also be used to interface with DrugEx, but to a more limited extent.

[mmahmedgithub]

Thanks so much Martin, for the detailed response. I’ll give this a try and see how it goes. In the meantime, it might be valuable to create a tutorial at this level of complexity, I imagine advanced users would be very interested. DrugEx is really unique, thanks again.